Improvement And Implementation Of Several Methods Based On Convolutional Neural Networks

Due to the progress in science as well as the introduction and implementation of new technologies such as 5G communication,Industry 4.0 and Internet of Things,the amount of data information has also experienced a marked increase.The advent of the big data era brings more problems to data processing.As regards big data,how to leverage massive data to better service human beings is a new subject.Deep learning has gradually risen to prominence.As a branch of artificial intelligence.it is mainly applied to solving and classifying complicated problems based on big data and boasts a strength of replacing manual feature acquisition with unsupervised or semi-supervised feature learning and hierarchical feature extraction algorithms.The main work and innovations of this paper include:1.Based on the neural network,in the face of a clear problem solving,its parameter setting and strategy mode lacks theoretical guidance.Based on the Let Net5 model,this paper compares the number of filters,pooling strategy,activation function and regularization strategy in the network model,and gives theoretical and practical data for the parameters and design of the convolutional neural network model.2.The convolution extraction part of traditional convolutional neural network is randomly generated convolution kernel for feature extraction.To ensure the diversity of feature extraction of network model,this paper presents a multi-feature extraction network model,and its multi-feature extraction structure is “Gabor feature + CNN feature + WT feature".Due to the multi-feature extraction of the network model,the data volume is increased.Therefore,two models based on different feature fusion algorithms are proposed.The two network models adopt PCA feature data fusion and multi-feature serial fusion algorithm.Finally,the feasibility of designing the network model through data set verification.3.The convolutional neural network obtains the classification result by linear feature mapping.Optimize its classifier,first pick the machine learning classification algorithm,and compare the random forest,GNB,SVM,linear discriminant classification algorithms in Caltech 101 dataset,and select the support vector machine and the improved based on the experimental results.The convolutional neural network is fused to design a hybrid model network model.4.The convolutional neural network will suffer from feature loss,limitation of feature information extraction and large data training model.The fusion neural network model combined with multi-feature extraction and capsule neural network are designed in this paper.The detailed information simplifies the feature dimension reduction part in the convolutional network,and uses the capsule neural network to improve the generalization ability of the hybrid network model,increase the spatial relationship judgment ability,make the advantages of the two network models complement each other,and improve the capsule network separately.Five different squashing functions are proposed,and the reconstructed network is compared experimentally to optimize the final hybrid network model.The final network model verifies the feasibility through the data set.